Edit Large YAML Prototypes with AI

AI configuration includes Model context window (tokens), defaulting to 50,000. Change it to the limit supported by your selected model and endpoint; selecting a different provider preserves the value. Setting a larger number does not increase a model's real capacity. Local models may use a smaller configured runtime context than their published maximum.

Vinyasa reserves 4,096 tokens for the response and 2,048 for protocol overhead. Input accounting conservatively treats each UTF-8 byte as one token rather than assuming that every language uses the same characters-per-token ratio. This deliberately underuses some models' windows. It is a safety estimate, not provider usage telemetry or an exact tokenizer.

Copilot does not send the full document or full conversation every turn. It starts with your request, element counts/types, a bounded page overview, and supported type names. The model may request schemas for up to eight types, search summaries in batches of twenty, and read up to twelve element IDs with selected properties.

Each request retains the original task, a short working plan, and the most recent inspection. Oversized fields are identified by name and length rather than silently presented as complete values. Only IDs actually returned by read inspections may be updated or removed. The final proposal contains add/update/remove operations; local validation merges updates into the original document, preserving untouched values. A small edit therefore does not require the AI to regenerate thousands of objects.

A generation allows at most ten model requests, with a 60-second timeout per request and a 4,096-token output cap. Requests and inspection batches are budget-checked. If a prompt, inspection, or valid result cannot fit, generation stops with an actionable error. There is no automatic provider fallback or retry loop.

This is not unlimited autonomous editing: the final edit list is capped at 200 operations, very long individual text values are omitted from reads above the field budget, and the agent cannot modify pages or workspace settings. Ask for a focused change to one component or screen, apply it after review, then continue with the next task. Editing an entire long text value is better done directly in YAML.

Prefer "Move the Save button on the billing screen 24 pixels right; preserve its ID and all text" over "Redesign every screen." Check the diff for removed objects, changed labels, and unintended coordinates. Discard proposals that miss the task. Cancel aborts current/future requests, but already-sent data and provider charges cannot be recalled.

When canvas changes make a draft stale, reload explicitly before applying. Back up the applied project through RVV and Git. For model setup and consent, start with AI Copilot.